CARIBOU: Computational AI Research Interface for Bioinformatics, Omics, and Unifying Agents

The growing gap between biological data generation and the availability of expert analysts motivates the development of AI systems capable of autonomously performing meaningful computational biology workflows. Here, we present CARIBOU (Computational AI Research Interface for Bioinformatics, Omics, and Unifying Agents), a multi-agent framework designed for practical deployment within institutional research computing environments. CARIBOU organizes specialized AI agents through researcher-modifiable blueprints that encode analytical roles, domain knowledge, and workflow guidance. All analyses are executed within reproducible computational environments compatible with Singularity/Apptainer-based high-performance computing (HPC) systems commonly used in academic research, while maintaining a persistent shared analytical state across multiple stages of analysis. This design enables CARIBOU to iteratively execute, troubleshoot, and refine bioinformatics workflows rather than simply generate static code. We evaluate CARIBOU across unit-task benchmarks, a metadata reconstruction challenge spanning six public datasets, and two end-to-end single-cell RNA-seq analyses using Allen Brain Atlas hippocampus and Tabula Sapiens large intestine datasets, alongside qualitative case studies demonstrating adaptive reasoning during analysis execution.